Three days of energy analysis, automated to 20 seconds

RDM's team were spending up to three days analysing energy data for a single supermarket location, when scaled across a 90-store chain, it would mean nearly two months of analyst time. Which just wasn’t feasible. In a four-week sprint, we automated the entire process, cutting the time per report to just 20 seconds. This not only enabled chain-wide reporting in under 30 minutes, but also laid the foundation for future AI-driven insights and scalability.

Walk through any supermarket on your weekly food shop, you'll pass dozens of refrigeration systems without giving them a second thought. These systems aren't like the fridge you or I have at home: they are a complex network of interconnected systems, often tied into the building's heating and air conditioning. They are extremely powerful and also extremely high-tech.

They also generate enormous amounts of data, which can be used to diagnose problems, often before they even happen. Which is great news. If you can process it quickly.

That's where RDM come in. They're the people keeping tabs on tens of thousands of refrigeration units across the country, catching problems before they turn into spoiled stock or emergency callouts.

The challenge

RDM have incredible expertise in building and maintaining refrigeration systems, and know their data inside out. But they had a problem: scale.

On average, a highly experienced Product Manager at RDM was spending up to three days analysing energy usage for just one supermarket location. This was mainly because the task involved merging 91 CSV files, totalling 104,000 rows and 150MB of data… manually.

With 90 stores on the contract, that's nearly two months of analyst time for a single reporting cycle.

RDM realised that if they could automate large parts of this process, not only would it save them time, but they would be able to forecast energy usage far more efficiently, and even look at using AI to predict component failures before they happen.

Our job was to make that automation real, quickly.

What we built

The amount of data we're dealing with is absolutely overwhelming if you're juggling it in spreadsheets but, for us, the challenge wasn't in the size of the data but in understanding enough about the current systems to be able to find the right processes to streamline and pain points to remove.

Over four weeks, we mapped their existing analysis processes, identified the bottlenecks, and replaced the manual CSV-merging workflow with a Python pipeline built around tools their team could actually use.

• In our first workshops, we took a deep dive into the kinds of data analysis tasks the team were doing on a regular basis, immediately proving value by showing that we could automate their time-consuming data aggregation and analysis processes.
• RDM have some great engineers but they're not data scientists, nor do they work in Python. We needed to make sure that the tools were useful with skillsets RDM had in house. To this end, we created configuration tools to automate running of new experiments and analysis. This meant the team were free to create new pipelines or test new ideas with minimal hassle.
• The ultimate plan would be for the things that we built to be fully integrated into everyday life at RDM, but we also needed to get the prototype in front of key stakeholders in the simplest way possible. To make what we build suitable for today and for the future, we built a containerised solution that could run anywhere from RDM's main servers to the CEO's laptop.

How we built it

There were two key considerations in deciding what to build and how to build it.

1. We wanted to demonstrate the immediate value of MLOps.
2. We wanted to design our tools to be easily and fully integrated into RDM's infrastructure.

With these considerations in mind, we broke the work into three stages.

• First, we replaced a manual, time-consuming process with an efficient, repeatable one. We converted the existing spreadsheet-driven analysis into a collection of tools built with Python and pandas.
• Next, we deployed machine learning pipelines that supported ongoing experimentation: Integrating our analysis tools into MLFlow allowed for quick experimentation and tracking of all our experiments.
• Finally, we packaged the solution in a lightweight, containerised format for portability and speed: This both allowed for experimentation on local machines by individual developers and offered a pathway for integration into RDM's core systems.

The results

The time required to generate the report for a single store was cut from three days to just twenty seconds. As a result, RDM could now run this vital report across all 90 stores in under 30 minutes.

What's now possible

The outcomes extended far beyond time savings:

• Scalability: Enabled consistent analysis across many locations.
• Customisability: Made it easy to adapt for different data sets and tasks.
• Future-readiness: Created a roadmap for integrating more advanced AI/ML capabilities, like automated failure prediction.
• Strategic potential: Set the stage for a company-wide MLOps workbench that could benefit all RDM clients.

In a short project, we were able to hit the ground running with actionable insights, automating what was taking months, and distilling into minutes. But, as with all midnight snacks, the fridge door's still open for more.